|Publication number||US7174415 B2|
|Application number||US 10/166,042|
|Publication date||Feb 6, 2007|
|Filing date||Jun 11, 2002|
|Priority date||Jun 11, 2001|
|Also published as||EP1267272A2, EP1267272A3, EP1267272B1, US20030012062|
|Publication number||10166042, 166042, US 7174415 B2, US 7174415B2, US-B2-7174415, US7174415 B2, US7174415B2|
|Inventors||Alon Ironi, Shachaf Zak|
|Original Assignee||Zoran Corporation|
|Export Citation||BiBTeX, EndNote, RefMan|
|Patent Citations (22), Non-Patent Citations (4), Referenced by (5), Classifications (15), Legal Events (7)|
|External Links: USPTO, USPTO Assignment, Espacenet|
The present application claims the benefit of U.S. Provisional Appln. No. 60/296,733, filed Jun. 11, 2001.
This invention relates to parallel processing.
In order to enable video capability, CPU 110 must be provided with software to handle video and graphics. Special instructions (often termed multi-media extension instructions) are designed for CPU 110, and video/graphics instructions are executed in series with other instruction. In addition, the frequency of CPU 110 must be increased compared to a CPU in a system without video capability so as to meet the real time requirement. The increased clock frequency linearly increases power consumption.
An additional source of power loss relates to the data transfer across the input/output pins (interconnect) between the CPU 110 IC and the memory 140 IC. This kind of data traffic involves charge/discharge of a large capacitive load on the input/output buffers and therefore large power consumption. Quantitatively, the amount of power wasted over the input/output pins of memory IC 140 is given by:
P I =P ID +P IC
where PID is the power consumed over the data pins, and PIC is the power consumed over the control and address pins. Typically, the former is much lager than the latter, and
PID =C IO *V IO 2*0.5*B IO
For example for a typical DRAM memory IC, the wasted power can be approximated as follows:
In another prior art system 200 of a cellular phone with video capability illustrated in
Both systems 300 and 320 have an increased IC count compared to systems 100 and 200 and therefore an increased size and cost. Systems 300 and 320 are also not power efficient because data has to be moved between CPU 310 and memory 140, between CPU 310 and DSP 385 or application specific CPU 380, and between DSP 385 or application specific CPU 380 and memory 395 or 390, consuming a lot of power. The data transfers between DSP 385 or application specific CPU 380 and memory 395 or 390 is typically (although not necessarily) the highest traffic of the data transfers listed above for systems 300 and 320, and therefore the most wasteful in power because video compression/decompression algorithms require multiple accesses to the data
There are also related art systems which include a processor embedded in CPU memory.
U.S. Pat. No. 6,026,478 to Dowling describes a VLIW (very large instruction word) processor that is connected to an embedded DRAM VLIW extension processor, that also functions as the DRAM of the VLIW CPU. U.S. Pat. No. 6,026,478 partitions the allocation of tasks between the CPU and the embedded DRAM processor at the instruction level.
As a result of splitting the program at the instruction level, the amount of data (where “data” includes instructions) exchanges between the CPU and the embedded DRAM processor is decreased by a certain amount with respect to a regular CPU-DRAM paradigm. The disadvantage of the system is in the required complexity. For example, the embedded memory processing unit needs to sense the stream of instructions that are executed at the same time by the CPU, and the embedded memory processing unit needs to share the instruction caching of the CPU. In addition, there are still significant data exchanges between the CPU and the DRAM in order to synchronize the execution at the instruction level.
U.S. Pat. No. 5,396,641 to Iobst et al. describes a single instruction multiple data (SIMD) and multiple instruction multiple data (MIMD) processors integrated with DRAM. Each type of processor has an external DRAM interface (i.e., the DRAM can be accessed as a common DRAM). However, there are extra control lines for operating the embedded processor from an external “host”.
Moreover the invention disclosed in U.S. Pat. No. 5,396,641 does not support simultaneous internal processing and external data transfers. An internal computation cycle can only take place instead of a memory access cycle. This approach makes it impossible to use the embedded DRAM processor at the same time as the CPU uses the embedded DRAM as its main memory.
U.S. Pat. No. 5,678,021 to Pawate et al. discloses a smart memory that includes a data storage and a processing core for executing instructions stored in the data storage area. Externally, the smart memory is directly accessible as a standard memory device. However the smart memory does not support simultaneous internal processing and external data transfers. An internal computation cycle can only take place instead of a memory access cycle. This approach makes it impossible to use the processing core at the same time as the CPU uses the data storage as its main memory.
What is needed in the art is a logic embedded memory where the memory can be accessed simultaneously by an embedded ASSPU and by an main processing unit.
The invention provides for a specialized memory device comprising:
a memory module, which includes at least two separate memory banks, including at least one specialized memory bank and at least one regular memory bank;
an application specific signal processing unit ASSPU which is embedded on the same die with said memory module, and bas a read/write access to said at least one specialized memory bank via an internal bus, said ASSPU configured to execute at least one pre-determined task;
a bus interface coupling an external main processing unit to said memory module, said bus interface enables said main processing unit to access said at least one specialized memory bank in respect of data related to at least one of said at least one predetermined tasks executable by said ASSPU, and enables said main processing unit to access said at least one regular memory bank; and
a traffic management unit TMU which allows access to at least one of said at least one specialized bank by said ASSPU simultaneously with access to at least one of said at least one regular memory bank by said main processing unit.
The invention further provides for a method for sharing memory in a memory module between a main processing unit and an application specific signal processing unit ASSPU which is embedded on the same die as the memory module, wherein, the memory module includes at least one bank shared by the main processing unit and the ASSPU and at least one regular bank accessible by the main processing unit, the method comprising:
determining whether the main processing unit desires to access a bank of the memory module which is shared by the main processing unit and by the ASSPU, in respect of data related to a predetermined task executable by the ASSPU; and
if said determination is that the main processing unit desires to access a non-shared (regular) bank of the memory module, allowing access to said non-shared bank by the main processing unit and allowing the ASSPU to access said shared bank simultaneously.
Yet further, the invention provides for a specialized memory device comprising:
a memory module, which includes at least one specialized memory bank;
an application specific signal processing unit ASSPU, which is embedded on the same die with said memory module, and has a read/write access to said at least one specialized memory bank via an internal bus, said ASSPU configured to execute at least one pre-determined task;
a bus interface coupling a main processing unit to said memory module, said bus interface enables said main processing unit to access at least one register of said ASSPU mapped to at least one address of said at least one specialized bank;
a traffic management unit TMU which allows access by said main processing unit to at least one of said at least one registers and allows access by said ASSPU to at least one of said at least one specialized memory bank which includes said at least one of said at least one addresses, simultaneously.
Still further, the invention provides for a method for sharing memory on a memory module between a main processing unit and an application specific signal processing unit ASSPU which is embedded on the same die as the memory module and is configured to execute at least one pre-determined task, the memory module including at least one specialized memory bank, the method comprising:
determining whether the main processing unit desires to access one of at least one predetermined addresses in one of said at least one specialized memory bank;
if said determination is that the main processing unit desires to access one of said at least one predetermined address in one of said at least one specialized banks, allowing access to at least one register of said ASSPU which is mapped to said one address by the main processing unit; and
during said main processing unit access, allowing said ASSPU to access said one specialized bank which includes said one address, if said ASSPU so desires.
The invention provides for a method for sharing memory on a memory module between a main processing unit and an application specific signal processing unit ASSPU which is embedded on the same die as the memory module and is configured to execute at least one pre-determined task, the memory module including at least one specialized memory bank, the method comprising:
determining whether the main processing unit desires to access one of at least one specialized memory bank;
if said determination is that the main processing unit desires to access one of said at least one specialized banks, allowing the main processing unit to access indirectly said one specialized bank; and
during said main processing unit access, allowing said ASSPU to access directly said one specialized bank if said ASSPU so desires.
In order to understand the invention and to see how it may be carried out in practice, a preferred embodiment will now be described, by way of non-limiting example only, with reference to the accompanying drawings, in which:
Any embodiment applying the invention includes a main processing unit IC perhaps with other peripheral processing and interface and communication modules on the same IC, and a main memory IC with an embedded ASSPU. The main processing unit can be for example, a core with two processors (CPU/DSP), a CPU or a DSP. The main memory can be for example DRAM, SRAM or a combination that mixes DRAM and SRAM. In the description below for ease of explanation, it is assumed that the main processing unit is a CPU and the to memory is DRAM. It is also assured in many of the embodiments described below that the processed data is output to a color LCD but other outputs may be used in alternative preferred embodiments.
Preferably, the ASSPU handles one or more system tasks (i.e. functions) which are memory intensive while the CPU handles other tasks. Data exchanges between the CPU and the main memory are therefore reduced, lowering power consumption. In addition, neither an extra IC is added, nor its accompanying added cost and size. Typically but not necessarily, tasks are at the level of a sub-application (for example the frame level), rather than at the instruction level or at the level of a group of instructions. Functions that are allocated to the ASSPU are generally excluded or eliminated from the program of the CPU. An example of such a function can be compressing an image, decompressing an image, converting an image from one color space to another, etc.
Note that if the allocation is at a level other than the instruction level or group of instructions level, the synchronization between the ASSPU and CPU may be simplified compared to an allocation at the instruction level, for example because there is no need for the ASSPU to share the instruction caching of the CPU.
In one preferred embodiment of the present invention, the physical interface between the CPU IC and the DRAM IC (including the ASSPU) conforms with the industry standard of physical interfaces between CPU IC and a regular DRAM IC (without the ASSPU). In another preferred embodiment of the invention, dedicated control and/or status lines are added. For example, added pins can be used for an interrupt signal to the CPU, and/or for a signal indicating that the access is related to the tasks performed by the ASSPU rather than being a regular memory access (in embodiments where the address of the access is not used for distinguishing the type of access).
A special bus 550 connects ASSPU 505 with specialized bank 520. In the description below and appended claims, the terms “bus” and “bus interface” should be construed to include a plurality of “buses” and “bus interfaces”, respectively. Typically but not necessarily, special bus 550 is wider than the other buses 580 inside logic embedded memory 440 to allow ASSPU 505 maximum utilization of specialized bank 520 and operation at a low frequency.
ASSPU 505 processes the image/video data, for example by compressing, decompressing, filtering etc. Results are temporarily stored in specialized bank 520.
Traffic management unit (TMU) 510 allows CPU 410 and ASSPU 505 to share access to specialized bank 520. CPU 410 typically although not necessarily does not access specialized bank 520 except for feeding ASSPU 505 with new data or acquiring results of processing by ASSPU 505. TMU 510 distinguishes between video and/or graphics data from other data and therefore CPU 410 access to specialized bank 520 rather than regular banks 530 based preferably on either the address range exhibited on the memory address bus of CPU 410 or special dedicated signals within the interface between CPU 410 and embedded memory 440 (for example connected to control logic 540). For example, the dedicated signals can include a signal set high by CPU 410 to indicate that the data is video/graphics data and set low by CPU 410 to indicate other (general purpose) data.
In one preferred embodiment based on
In case of a CPU write, TMU 510 buffers the written data, while signaling ASSPU 505 to release access of specialized bank 520. Once specialized bank 520 is accessible to CPU 410, TMU 510 emulates a CPU write cycle, starting with the data that was temporarily buffered. Similarly, if CPU 410 is first in accessing specialized bank 520, TMU 410 prevents access by ASSPU 505.
In certain cases, conflict can be minimized or eliminated through synchronization of CPU 410 with respect to the specific application. For example, in case of video decompression, CPU 410 may suite the compressed data to a pre-defined buffer in memory 440, notify ASSPU 505 of that via a command data word that is delivered across the external memory interface, and write the next load of compressed data to another buffer, notify ASSPU 505 and so forth. This scheme of double or multiple buffers uses some extra memory space, but allows for a safe, sustained arbitration of memory portions to CPU 410 and ASSPU 505 without creating conflict situations.
Another preferred embodiment of the internal configuration of embedded memory 440 is illustrated in
TMU 1210 distinguishes between access by CPU 410 to regular banks 530 or specialized bank 520, based either on the address or by dedicated signals connected to the control logic as was discussed above.
TMU 1210 controls a de-multiplexer 1260 and a multiplexer 1270. De-multiplexer 1260 has one input from data input buffer 540 and two outputs, data bus 580 or FIFO input register 1230, depending on a control signal from TMU 1210. Multiplexer 1270 has two inputs (from data bus 580 or from FIFO output register 1240) and one output to data output buffer 545. A control signal from TMU 1210 selects which of the two inputs will be transferred to data output buffer 545. Note that in the preferred embodiment illustrated in
The steps of an example of a video decompression by system 400 will now be explained with reference to
Steps 610, 612, 614 and 616 are repeated for each subsequent segment in the frame. In step 618, CPU 410 completes the waiting of the current frame to memory. In step 620, ASSPU 505 asserts a flag noticing CPU 410 that the last segment is ready. In step 622, CPU 410 starts writing the next compressed frame to memory. Note that the writing of the next compressed frame can occur before reading the last ready segment of the previous frame. In step 624, CPU 410 notifies ASSPU 505 that a new compressed frame has started to be inputted into memory. In step 626, ASSPU 505 begins decompressing the new frame and the process begins anew.
In an alternative functional flow, there are no or less notifications between CPU 410 and ASSPU 505. For example, assuming an internal structure as illustrated in
The process of CPU 710 writing data to logic embedded memory 740 is similar to the process illustrated in
Refer to FIG, 8, which illustrates the processing by a graphics accelerator for a cellular telephone configured as in
Recall that in a cellular phone system without the present invention, the processing of the high level graphics command to data exchanges are done within the CPU, or by an application specific signal processor or DSP IC that bridges between the CPU and the LCD. In both cases, the data exchange between the graphics processing logic and memory is intensive and consumes a lot of power. The cellular telephone of the present invention has the graphics processing logic as part of the logic embedded memory IC. This replaces “external” data exchanges between the CPU IC and memory IC with “internal” data exchanges between the ASSPU and part of the memory array on the same IC. Being internal, and at a lower frequency, these internal exchanges consume less power. However, in preferred embodiments where CPU 410 reads the resultant processed data from memory 440 in order to pass the data to LCD 120, the data passes through input/output, and power is consumed. As an example, LCD 120 can require updated data 60 times per second. Therefore the preferred embodiment illustrated in
The logic embedded memory of the current invention can be used for other applications as presented in the following preferred embodiments.
CBEP 910 accesses regular banks 530 for regular (general purpose) data exchanges, such as accesses to the On-Screen Display (OSD) memory, graphics frame buffer, etc, for example when the CBEP 910 cache has “misses”.
In some embodiments where the system of
When the incoming signal is a compressed digital stream, this feature mainly involves the operation of file formatting and recording. However, when the source is legacy analog TV signal (NTSC/PAL/SECAM) the operation involves (digitization and) high performance MPEG2 compression.
Typically although not necessarily, digital TV set-top box systems have an STB integrated processor that takes care of parsing the incoming bit stream, decompressing the video and sending it for display, decompressing the audio and playing it, and other functions such as graphical user interface, decryption management, electronic programming guide applications, etc. The base-band processor typically although not necessarily uses a large amount of DRAM—128 Mbit and more.
This processor also takes care of recording video and audio when the source is digital and compressed. In order to support the storage of video sourced out of an analog signal, or even to support bit-rate conversion of video sourced out of the digital stream, video compression is needed. The compression standard used in set-top box systems is MPEG2. This standard is very much a memory-intensive and computationally power intensive process.
The addition of logic embedded memory 440 enables video compression. This solution is advantageous compared to the prior art solution of adding a dedicated application specific signal processor plus internal or external DRAM, and thereby increasing IC count, cost, power, and if the DRAM is external then also DRAM redundancy. The solution is also advantageous compared to a second prior art solution of enhancing the STB integrated processor with an MPEG2 compression core and increasing DRAM sufficiently for the MPEG2 compression process, thereby creating a bottleneck of data throughput between the processor and the DRAM. The use of two DRAM IC's in the second prior art solution, one for compression and the other for other tasks with separate busses connecting the processor to the DRAM ICs would solve the bottleneck but at the cost of expense and DRAM redundancy.
In the case of recording video sourced in an analog signal 1116, the analog data is digitized and decoded by a video decoder 1122, transferred through set top box STB integrated processor 1110 to logic embedded memory 440. The internal embedded ASSPU 505 compresses the video data. STB integrated processor 1110 transfers the data from logic embedded memory 440 to hard-disk controller 1124, and from there to the hard-disk 1126. In an alternative preferred embodiment, video decoder 1122 can be embedded within STB integrated processor 1110, mutatis mutandis.
As mentioned above, even if the video source is digital, i.e. the video is input compressed, there may be a need for MPEG2 compression, e.g. in order to reduce the bit-rate of the incoming video so as to make it more compact before its storage. This requires nearly a complete MPEG2 process, utilizing logic embedded memory 440 of the present invention.
Alternatively, the compressed video from a logic embedded memory can directly output to a display such as a TV.
It should be noted that although the design of high performance, dense logic processing functions requires a different semiconductor process than the design of high performance dense memory, the cost increment of manufacturing a memory with an ASSPU using a memory-optimized semi-conductor process compared to a memory without an ASSPU is relatively small if the amount of logic is significantly smaller an the amount of memory. The power savings however are relatively large due to the data exchanges being internal between the ASSPU and the specialized bank. To further illustrate the power savings the following example is presented:
P 2 =C in *V in 2*0.5*B where
In addition, if the ASSPU 505 is embedded in DRAM memory and has a wide bus, the invention may allow a saving of SRAM, by replacing SRAM with DRAM, thereby reducing cost because DRAM is more dense than SRAM.
The present invention is especially advantageous under one or more of the following industrial conditions:
1) The system is required to exhibit low power consumption, low IC count (and/or small physical size), and low cost;
2) The semi-conductor area required for implementing the digital signal, processing function (e.g., video compression, decompression, etc.) is much smaller than the semi-conductor area required for implementing the overall memory required for the operation of the CPU and the operation of the digital signal processing function; and/or
3) The required digital signal processing function does not need heavy data exchange with the CPU, relative to the data exchange that the signal processing function requires with memory.
It will also be understood that the system according to the invention may be a suitably programmed computer. Likewise, the invention contemplates a computer program being readable by a computer for executing the method of the invention. The invention further contemplates a machine-readable memory tangibly embodying a program of instructions executable by the machine for executing the method of the invention.
While the invention has been described with respect to a limited number of embodiments, it will be appreciated that many variations, modifications and other applications of the invention may be made.
|Cited Patent||Filing date||Publication date||Applicant||Title|
|US5396641||Jun 8, 1993||Mar 7, 1995||Iobst; Kenneth W.||Reconfigurable memory processor|
|US5678021||Oct 17, 1994||Oct 14, 1997||Texas Instruments Incorporated||Apparatus and method for a memory unit with a processor integrated therein|
|US5761719 *||Jun 6, 1995||Jun 2, 1998||International Business Machines Corporation||On-chip memory map for processor cache macro|
|US5815167 *||Jun 27, 1996||Sep 29, 1998||Intel Corporation||Method and apparatus for providing concurrent access by a plurality of agents to a shared memory|
|US5854638||Feb 2, 1996||Dec 29, 1998||Opti Inc.||Unified memory architecture with parallel access by host and video controller|
|US5870625 *||Dec 11, 1995||Feb 9, 1999||Industrial Technology Research Institute||Non-blocking memory write/read mechanism by combining two pending commands write and read in buffer and executing the combined command in advance of other pending command|
|US5941968 *||Apr 14, 1997||Aug 24, 1999||Advanced Micro Devices, Inc.||Computer system for concurrent data transferring between graphic controller and unified system memory and between CPU and expansion bus device|
|US6000027||Aug 25, 1992||Dec 7, 1999||Texas Instruments Incorporated||Method and apparatus for improved graphics/image processing using a processor and a memory|
|US6020903||Sep 22, 1997||Feb 1, 2000||Samsung Electronics Co., Ltd.||Method and apparatus for using video memory as system memory|
|US6026478||Dec 23, 1997||Feb 15, 2000||Micron Technology, Inc.||Split embedded DRAM processor|
|US6070002||Sep 13, 1996||May 30, 2000||Silicon Graphics, Inc.||System software for use in a graphics computer system having a shared system memory|
|US6170048 *||Aug 11, 1999||Jan 2, 2001||Texas Instruments Incorporated||PC circuits, systems and methods|
|US6185704||Apr 9, 1998||Feb 6, 2001||Texas Instruments Incorporated||System signaling schemes for processor and memory module|
|US6330646 *||Jan 8, 1999||Dec 11, 2001||Intel Corporation||Arbitration mechanism for a computer system having a unified memory architecture|
|US6381671||Jun 29, 1999||Apr 30, 2002||Hitachi, Ltd.||Semiconductor integrated circuit and data processing system|
|US6584588||Oct 30, 2000||Jun 24, 2003||Texas Instruments Incorporated||System signalling schemes for processor & memory module|
|EP0446721A2||Mar 1, 1991||Sep 18, 1991||Texas Instruments Incorporated||Distributed processing memory|
|JP2000021160A||Title not available|
|JPH05342140A||Title not available|
|JPH06208632A||Title not available|
|JPH08221313A||Title not available|
|JPH10289151A||Title not available|
|1||European Search Report for EP Application No. 02012940, dated May 22, 2006.|
|2||*||IEEE 100: The Authoritative Dictionary of IEEE Standards Terms, 7<SUP>th </SUP>Ed., IEEE Press, Dec. 2000, pp. 123 and 949.|
|3||*||Intel 82810/82810-DC100 Graphics and Memory Controller Hub (GMCH), Intel Corporation, 1999, pp. 1-17.|
|4||Murakami, K., "Columubus's Egg," PEMM Combination of CPU and DSP, Monthly Electronics Magazine, vol. 44, 4th Ed., Apr. 1, 1999, pp. 77-81.|
|Citing Patent||Filing date||Publication date||Applicant||Title|
|US7353319 *||Jun 2, 2005||Apr 1, 2008||Qualcomm Incorporated||Method and apparatus for segregating shared and non-shared data in cache memory banks|
|US7821849 *||Oct 26, 2010||Infineon Technologies Ag||Configurable embedded processor|
|US8270231 *||Sep 18, 2012||Infineon Technologies Ag||Configurable embedded processor|
|US20060277356 *||Jun 2, 2005||Dec 7, 2006||Speier Thomas P||Method and apparatus for managing cache memory accesses|
|US20080195835 *||Feb 8, 2008||Aug 14, 2008||Infineon Technologies Ag||Configurable embedded processor|
|U.S. Classification||711/5, 365/230.03|
|International Classification||G06F13/16, G06F12/06, G06F13/28, G11C11/401, G11C7/00, G06F15/78, G06F12/00|
|Cooperative Classification||G06F15/7821, G06F13/1647, G06F9/3877|
|European Classification||G06F9/38S, G06F13/16A6, G06F15/78P1D|
|Nov 25, 2002||AS||Assignment|
Owner name: EMBLAZE SEMICONDUCTOR LTD., ISRAEL
Free format text: ASSIGNMENT OF ASSIGNORS INTEREST;ASSIGNORS:IRONI, ALON;ZAK, SHACHAF;REEL/FRAME:013520/0004;SIGNING DATES FROM 20020611 TO 20020616
|Dec 16, 2004||AS||Assignment|
Owner name: ZORAN MICROELECTRONICS MOBILE (1995) LTD., ISRAEL
Free format text: CHANGE OF NAME;ASSIGNOR:EMBLAZE SEMICONDUCTOR LTD.;REEL/FRAME:015464/0026
Effective date: 20040708
|Dec 26, 2004||AS||Assignment|
Owner name: ZORAN MICROELECTRONICS LTD., ISRAEL
Free format text: ASSIGNMENT OF ASSIGNORS INTEREST;ASSIGNOR:ZORAN MICROELECTRONICS MOBILE (1995) LTD.;REEL/FRAME:015485/0931
Effective date: 20041221
|Aug 6, 2010||FPAY||Fee payment|
Year of fee payment: 4
|Jun 11, 2014||AS||Assignment|
Owner name: QUALCOMM TECHNOLOGIES, INC., CALIFORNIA
Free format text: ASSIGNMENT OF ASSIGNORS INTEREST;ASSIGNOR:CSR ISRAEL LTD.;REEL/FRAME:033074/0476
Effective date: 20140608
|Jun 16, 2014||AS||Assignment|
Owner name: CSR ISRAEL LTD, ISRAEL
Free format text: CHANGE OF NAME;ASSIGNOR:ZORAN MICROELECTRONICS LTD;REEL/FRAME:033113/0766
Effective date: 20130814
|Jul 25, 2014||FPAY||Fee payment|
Year of fee payment: 8